Temporal awareness: due_at, reminded_at, and proactive surfacing

[kovtcharov]

Summary

Implement temporal awareness in the memory system — the agent knows what time it is, what's coming up, and proactively surfaces time-sensitive items.

Spec: docs/spec/agent-memory-architecture.md → Temporal sections throughout

Depends on: #542 (MemoryStore), #543 (MemoryMixin)

Features

Schema

• due_at column on knowledge — ISO 8601 future date
• reminded_at column — when agent last surfaced this to user
• Partial index on due_at WHERE due_at IS NOT NULL

MemoryStore

• get_upcoming(within_days=7) — returns items due soon or overdue
• Query: due_at <= future AND (reminded_at IS NULL OR reminded_at < due_at)
• All temporal queries use Python-computed boundaries (SQLite doesn't handle TZ offsets)

System Prompt

• Current date/time always injected
• Upcoming/overdue section with [DUE date] and [OVERDUE date] labels
• Instruction to call update_memory(reminded_at="now") after mentioning

Tools

• remember(due_at=...) — store time-sensitive items
• update_memory(reminded_at="now") — mark as mentioned
• Validated via datetime.fromisoformat()

Reminder Lifecycle

1. User tells agent about future event → stored with due_at
2. Agent interacts near due date → sees in system prompt → mentions it
3. Agent marks reminded_at → won't nag
4. If due_at passes and was only reminded before → resurfaces as overdue

Acceptance Criteria

• due_at stored and validated correctly
• get_upcoming returns correct items within time window
• Overdue items resurface after reminded_at < due_at
• System prompt includes temporal section
• Current time injected in system prompt
• LLM can set reminded_at via update_memory